Rank 7 Solution for JanathaHack: Recommendation System
The client has provided you with history of last 10 challenges the user has solved, and you need to predict which might be the next 3 challenges the user might be interested to solve. Apply your data science skills to help the client make a big mark in their user engagements/revenue.
- Taken a Text Generation Approach and tried training with each sequence combination Ex: for sequence (1,2,3,4,5,6...10) I have taken taken combinations of (1,2-->3), (1,2,3-->4), (1,2,3,4-->5)...
- This Approach gave me a score of 0.05 map which is very less to start with.
- After Lot of effort in LSTM(which i couldnt improve) i searched how else can we solve this problem and found that turicreate API works very well with out any actual training. It is suggested to use this API for quick and better solutions of recommendation Link to Turicreate
- Using Turicreate Itemsimilarity recommender gave me a score of 0.2011 map.
- Not satisfied with the 0.201 score and again tried various approaches of input data.
- Finally FinalLSTM.py approach worked , check it Here
- Python for programming
- pandas and numpy libraries for methodology
- Tensorflow and keras library for the model
- matplotlib and seaborn was used for plotting and analyzing the data
Final Approach is taken from McKinsey Analytics Online Hackathon- Recommendation Design Here
Rank: 7th on public LB and 7th on private LB
Link